Joint Latent Dirichlet Allocation for Social Tags
Social tags, serving as a textual source of simple but useful semantic metadata to reflect the user preference or describe the web objects, has been widely used in many applications. However, social tags have several unique characteristics, i.e., sparseness and data coupling (i.e. non-IIDness), which makes existing text analysis methods such as LDA not directly applicable. In this article, we propose a new generative algorithm for social tag analysis named Joint Latent Dirichlet Allocation, which models the generation of tags based on both the users and the objects, and thus accounts for the coupling relationships among social tags. The model introduces two latent factors that jointly influence tag generation: the user’s latent interest factor and the object’s latent topic factor, formulated as user-topic distribution matrix and object-topic distribution matrix, respectively. A Gibbs sampling approach is adopted to simultaneously infer the above two matrices as well as a topic-word distribution matrix. Experimental results on four social tagging data sets have shown that our model is able to capture more reasonable topics and achieves better performance than five state-of-the-art topic models in terms of the widely used point-wise mutual information (PMI) metric. In addition, we analyze the learnt topics showing that our model recovers more themes from social tags while LDA may lead the topic vanishing problems, and demonstrate its advantages in the social recommendation by evaluating the retrieval results with Mean Reciprocal Rank (MRR) metric. Finally, we explore the joint procedure of our model in depth to show the non-IID characteristic of social tagging process. Index Terms—Topic models, social tags, Non-IID learning, Joint Latent Dirichlet Allocation, structure mining
With the explosion of online media, many websites such as Flickr, Delicious and CiteULike provide the tagging services to annotate media objects (e.g., image, video, music and web page) in the forms of short text. These data, termed as social tags, present us new opportunities as well as challenges to better understand the multimedia objects or user behavior patterns on the internet, and have been widely used in many applications including multimedia retrieval image analysis recommender system, and user modeling. However, although general text analysis methods such as LDA or its variants have shown promise for social tag analysis, these models fail to consider the following two unique characteristics of social tags. First, the amount of tags associated with one object marked by one user is sparse, usually ranging from 1 to 10, which makes existing topic models hard to learn reliable features from this single source. Second and the most importantly, social tags are not independent and identical distributed (non-IID) and are generated based on coupling the user interests and the inherent characteristics of the object. For example, suppose there is an image about sunset at the beach. Based on users’ preferences, the image may be tagged as fsunset, goldeng or fbeach, sandg. Most existing methods for text analysis tend to ignore the contribution of either the user’s interest or object’s inherent characteristics in tag generation.
In this paper we explore a novel approach for social tag analysis taking full consideration of its non-IID characteristic. Fig. 1 provides an illustration of the tagging graph and the generative procedure of social tags1. Note that Doc here refers to different kinds of web objects. Besides, different from multi-source models, we do not utilize the content of objects and both the users and the objects are modeled by latent variables, each object may be assigned different tags by different users due to their preferences and each user (e.g. User 3) may give different tags for different objects according to the inherent characteristics of the objects, i.e., the generation of the social tags depends on both the users and the objects. Based on this analysis, the generation of social tags is illustrated. Given a user, an object and the tags, we represent the user with a latent interest distribution and the object with a latent topic distribution. Then to generate a tag, a joint topic is first formed and a tag is subsequently sampled based the joint topic. We represent the above procedure with a generative model, i.e., Joint Latent Dirichlet Allocation (JLDA) model, which is an extension of Latent Dirichlet Allocation (LDA) model. Different from LDA, JLDA simultaneously accounts for the influence of the users and the objects on tag generation, which is expected to be more suitable for social tags analysis considering the non-IID characteristics. The advantage could be intuitively interpreted with, we find that in most cases, it is difficult to arise prominent topics due to the sparsity of tags for an object or one user. For example, if only considering the tags “mytilus”, “ocean” and “altantic” of Doc 3 or the tags “adsorption”, “environment” and “water” of User 2, it is difficult to interpret the actual meaning of these words regarding to Doc 3 and User 2. However, if we know User 3 is biology and environmentalist from his other tags and Doc 1 is referring to the environmental problems from the tags of other users, we may better understand Doc 3 and User 2. This suggests that it is helpful to understand the sematic meanings of social tags collaboratively with the users and the objects.
In this paper, we proposed a Joint Latent Dirichlet Allocation method for modeling social tags. By considering the non-IID characteristic of social tags, our model introduces two factors, the user interest factor and the object latent topic factor, to jointly affect the generative procedure of tags. As our model utilizes the collaborative information among the users and the objects to extract more explicit information from tags, experiments on four publicly available data sets have demonstrated the advantages of our models compared with other five topic models in terms of PMI scores. Besides, we analyze the learnt topics in Delicious data set and show that our model recovers more thematically topics from social tags while LDA may lead to vanishing topics by considering these social data in only one perspective. Following that, we further show the advantages of JLDA in the social recommendation application. Finally, we explore the joint procedure of social tags at the topic level and explain that our model is more interpretable than LDA in the generative procedure of social tags.
 X. Li, C. G. Snoek, and M. Worring, “Learning social tag relevance by neighbor voting,” Multimedia, IEEE Transactions on, vol. 11, no. 7, pp. 1310–1322, 2009.
 M. Levy and M. Sandler, “Music information retrieval using social tags and audio,” Multimedia, IEEE Transactions on, vol. 11, no. 3, pp. 383–395, 2009.
 L. Chen, D. Xu, I. W. Tsang, and J. Luo, “Tag-based image retrieval improved by augmented features and group-based refinement,” Multimedia, IEEE Transactions on, vol. 14, no. 4, pp. 1057–1067, 2012.
 Y. Mao, H. Shen, and C. Sun, “A probabilistic topic model with social tags for query reformulation in informational search,” in Advanced Data Mining and Applications, 2011.
 L. Wu, R. Jin, and A. Jain, “Tag completion for image retrieval,” PAMI Pattern Analysis and Machine Intelligence, IEEE Transactions on, 2013.
 Z. Xu, Y. Zhang, and L. Cao, “Social image analysis from a non-iid perspective,” Multimedia, IEEE Transactions on, vol. 16, no. 7, pp. 1986–1998, 2014.
 L.-J. Li, R. Socher, and L. Fei-Fei, “Towards total scene understanding: Classification, annotation and segmentation in an automatic framework,” in Computer Vision and Pattern Recognition (CVPR), IEEE Conference on, 2009, pp. 2036–2043.
 R. Socher and L. Fei-Fei, “Connecting modalities: Semisupervised segmentation and annotation of images using unaligned text corpora,” in Computer Vision and Pattern Recognition (CVPR), IEEE Conference on, 2010, pp. 966–973.
 Y. t. Song, “Automatic tag recommendation algorithms for social recommender systems,” ACM Transactions on the Web, 2011.
 F. t. Abel, “Cross-system user modeling and personalization on the social web,” User Modeling and User- Adapted Interaction, 2013.
 D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent dirichlet allocation,” JMLR the Journal of machine Learning research, 2003.